The Austrian Academy of Science, with Mistral and Sail Reply, is releasing Apollo, described as the world's first advanced large language model for Ancient Greek. Trained on roughly 600 million historical Greek words, the free chatbot helps scholars restore damaged papyri by proposing likely missing words and passages while keeping humans in the loop.
A new arXiv paper examines whether language models can restore text lost to physical lacunae in damaged ancient manuscripts. It proposes scenarios mimicking real-world damage, matches model architectures to each scenario, and introduces decoding strategies to handle mismatches between lacuna boundaries and tokenization. Findings indicate restoration performance remains limited, a relevant caution for digital humanities practitioners.
A new arXiv paper examines whether language models can restore text lost to physical lacunae in damaged ancient manuscripts. It proposes scenarios mimicking real-world damage, matches model architectures to each scenario, and introduces decoding strategies to handle mismatches between lacuna boundaries and tokenization. Findings indicate restoration performance remains limited, a relevant caution for digital humanities practitioners.
Researchers are applying AI to decipher Linear A, the Bronze Age Minoan writing system that has resisted translation for a century. The approach may offer a new path for language isolates lacking bilingual anchors.
Researchers are applying AI to decipher Linear A, the Bronze Age Minoan writing system that has resisted translation for a century. The approach may offer a new path for language isolates lacking bilingual anchors.
Researchers introduce Stoicheia, a 405M-parameter character-level masked-diffusion encoder for Ancient Greek that factors input into five aligned planes (letters, boundaries, diacritics, capitalization, punctuation), enabling a single backbone to restore lacunae, re-segment, accentuate, and punctuate without task-specific retokenization. The model is pretrained on an open corpus of 380M words and released as eleven checkpoints with decontamination.
Researchers introduce Stoicheia, a 405M-parameter character-level masked-diffusion encoder for Ancient Greek that factors input into five aligned planes (letters, boundaries, diacritics, capitalization, punctuation), enabling a single backbone to restore lacunae, re-segment, accentuate, and punctuate without task-specific retokenization. The model is pretrained on an open corpus of 380M words and released as eleven checkpoints with decontamination.
Researchers have trained a neural network that translates digitized Akkadian cuneiform tablets into English, opening up hundreds of thousands of ancient texts to broader study. The system targets one of the earliest Semitic languages, which was written for over three millennia.
Researchers have trained a neural network that translates digitized Akkadian cuneiform tablets into English, opening up hundreds of thousands of ancient texts to broader study. The system targets one of the earliest Semitic languages, which was written for over three millennia.
Researchers introduced AlphaOracle, a deep learning framework inspired by human expert workflows, to decipher the 3,000 undeciphered oracle bone script characters. The system uses the largest digitized corpus and a multi-stage pipeline including rubbing parsing, radical-based morphological analysis, and contextual retrieval. This could accelerate understanding of ancient Chinese civilization.
Researchers introduced AlphaOracle, a deep learning framework inspired by human expert workflows, to decipher the 3,000 undeciphered oracle bone script characters. The system uses the largest digitized corpus and a multi-stage pipeline including rubbing parsing, radical-based morphological analysis, and contextual retrieval. This could accelerate understanding of ancient Chinese civilization.